Main Content

Define Variable-Size Data for Code Generation

For code generation, before using variables in operations or returning them as outputs, you must assign them a specific class, size, and complexity. Generally, after the initial assignment, you cannot reassign variable properties. Therefore, after assigning a fixed size to a variable or structure field, attempts to grow the variable or structure field might cause a compilation error. In these cases, you must explicitly define the data as variable-size by using one of these methods.

MethodSee

Assign the data from a variable-size matrix constructor such as:

Use a Matrix Constructor with Nonconstant Dimensions
Assign multiple, constant sizes to the same variable before using (reading) the variable.Assign Multiple Sizes to the Same Variable
Grow an array by using (end + 1) indexing.Grow an Array Using (end + 1) Indexing
Define all instances of a variable to be variable-size.Define Variable-Size Data Explicitly by Using coder.varsize

Note

The MATLAB Function block does not support code generation if one of the variables uses an alias type and is variable size. This limitation does not apply to input or output variables. For more information on defining variable-size variables in MATLAB Function blocks, see Create and Define MATLAB Function Block Variables and Declare Variable-Size MATLAB Function Block Variables.

Use a Matrix Constructor with Nonconstant Dimensions

You can define a variable-size matrix by using a constructor with nonconstant dimensions. For example:

function s = var_by_assign(u) %#codegen
y = ones(3,u);
s = numel(y);

If you are not using dynamic memory allocation, you must also add an assert statement to provide upper bounds for the dimensions. For example:

function s = var_by_assign(u) %#codegen
assert (u < 20);
y = ones(3,u);
s = numel(y);

Assign Multiple Sizes to the Same Variable

Before you use (read) a variable in your code, you can make it variable-size by assigning multiple, constant sizes to it. When the code generator uses static allocation on the stack, it infers the upper bounds from the largest size specified for each dimension. When you assign the same size to a given dimension across all assignments, the code generator assumes that the dimension is fixed at that size. The assignments can specify different shapes and sizes.

When the code generator uses dynamic memory allocation, it does not check for upper bounds. It assumes that the variable-size data is unbounded.

Inferring Upper Bounds from Multiple Definitions with Different Shapes

function s = var_by_multiassign(u) %#codegen
if (u > 0)
    y = ones(3,4,5);
else
    y = zeros(3,1);
end
s = numel(y);

When the code generator uses static allocation, it infers that y is a matrix with three dimensions:

  • The first dimension is fixed at size 3

  • The second dimension is variable-size with an upper bound of 4

  • The third dimension is variable-size with an upper bound of 5

When the code generator uses dynamic allocation, it analyzes the dimensions of y differently:

  • The first dimension is fixed at size 3.

  • The second and third dimensions are unbounded.

Grow an Array Using (end + 1) Indexing

In MATLAB® execution, you can grow any dimension of an array using end. For example, you can grow both dimensions of a two-dimensional matrix by assigning a value to the (end + 1) element of each dimension.

A = magic(5)
A =

    17    24     1     8    15
    23     5     7    14    16
     4     6    13    20    22
    10    12    19    21     3
    11    18    25     2     9
A(end + 1,end + 1) = 1
A =

    17    24     1     8    15     0
    23     5     7    14    16     0
     4     6    13    20    22     0
    10    12    19    21     3     0
    11    18    25     2     9     0
     0     0     0     0     0     1

To grow arrays using end in MATLAB code for code generation, you must adhere to these restrictions:

  • You can only use end to grow vectors. For example, code generation fails for this function because X is a matrix, not a vector.

    function X = foo
    X = [1 2; 3 4];
    X(end + 1,:) = 5;
    end
  • You can grow a vector only by assigning a value to the (end + 1) element. Assigning a value to subsequent elements, such as (end + 2), is not supported.

  • You can grow an empty 1-by-0 array by using (end + 1). Growing a 0-by-1 array using (end + 1) is not supported. You can grow a 0-by-0 array using (end + 1) only if you create the array using [].

  • Using (end + 1) to grow a structure array that has been defined using dot notation is not supported. For example, code generation fails for this MATLAB function.

    function growStruct
    s.field1 = 5;
    s.field2 = 2;
    s(end + 1) = s;
    end
    To grow a structure array using (end + 1), define the fields of the structure using the struct constructor.
    function growStruct
    s = struct("field1",5,"field2",2);
    s(end + 1) = s;
    end
    To learn more about growing arrays using (end + 1), see Generate Code for Growing Arrays and Cell Arrays with end + 1 Indexing (MATLAB Coder).

Define Variable-Size Data Explicitly by Using coder.varsize

To explicitly define variable-size data, use the function coder.varsize. Optionally, you can also specify which dimensions vary along with their upper bounds. For example:

  • Define B as a variable-size 2-dimensional array, where each dimension has an upper bound of 64.

    coder.varsize('B', [64 64]);

  • Define B as a variable-size array:

    coder.varsize('B');

    When you supply only the first argument, coder.varsize assumes that all dimensions of B can vary and that the upper bound is size(B).

If a MATLAB Function block output signal is variable-size, in the Property Inspector, you must specify that the signal is variable-size. You can specify the upper bounds or define the variable as unbounded (since R2023b). You do not have to use coder.varsize with the corresponding output variable inside the MATLAB Function block. However, if you specify upper bounds with coder.varsize, they must match the upper bounds in the Property Inspector.

Specify Which Dimensions Vary

You can use the function coder.varsize to specify which dimensions vary. For example, the following statement defines B as an array whose first dimension is fixed at 2, but whose second dimension can grow to a size of 16:

coder.varsize('B',[2, 16],[0 1])
.

The third argument specifies which dimensions vary. This argument must be a logical vector or a double vector containing only zeros and ones. Dimensions that correspond to zeros or false have fixed size. Dimensions that correspond to ones or true vary in size. coder.varsize usually treats dimensions of size 1 as fixed. See Define Variable-Size Matrices with Singleton Dimensions.

Allow a Variable to Grow After Defining Fixed Dimensions

Function var_by_if defines matrix Y with fixed 2-by-2 dimensions before the first use (where the statement Y = Y + u reads from Y). However, coder.varsize defines Y as a variable-size matrix, allowing it to change size based on decision logic in the else clause:

function Y = var_by_if(u) %#codegen
if (u > 0)
    Y = zeros(2,2);
    coder.varsize('Y');
    if (u < 10)
        Y = Y + u;
    end
else
    Y = zeros(5,5);
end

Without coder.varsize, the code generator infers Y to be a fixed-size, 2-by-2 matrix. It generates a size mismatch error.

Define Variable-Size Matrices with Singleton Dimensions

A singleton dimension is a dimension for which size(A,dim) = 1. Singleton dimensions are fixed in size when:

  • You specify a dimension with an upper bound of 1 in coder.varsize expressions.

    For example, in this function, Y behaves like a vector with one variable-size dimension:

    function Y = dim_singleton(u) %#codegen
    Y = [1 2];
    coder.varsize('Y', [1 10]);
    if (u > 0)
        Y = [Y 3];
    else
        Y = [Y u];
    end
    

  • You initialize variable-size data with singleton dimensions by using matrix constructor expressions or matrix functions.

    For example, in this function, X and Y behave like vectors where only their second dimensions are variable-size.

    function [X,Y] = dim_singleton_vects(u) %#codegen
    Y = ones(1,3);
    X = [1 4];
    coder.varsize('Y','X');
    if (u > 0)
        Y = [Y u];
    else
        X = [X u];
    end

You can override this behavior by using coder.varsize to specify explicitly that singleton dimensions vary. For example:

function Y = dim_singleton_vary(u) %#codegen
Y = [1 2];
coder.varsize('Y', [1 10], [1 1]);
if (u > 0)
    Y = [Y Y+u];
else
    Y = [Y Y*u];
end

In this example, the third argument of coder.varsize is a vector of ones, indicating that each dimension of Y varies in size.

Define Variable-Size Structure Fields

To define structure fields as variable-size arrays, use a colon (:) as the index expression. The colon (:) indicates that all elements of the array are variable-size. For example:

function y=struct_example() %#codegen

d = struct('values', zeros(1,0), 'color', 0);
data = repmat(d, [3 3]);
coder.varsize('data(:).values');

for i = 1:numel(data)
    data(i).color = rand-0.5;
    data(i).values = 1:i;
end

y = 0;
for i = 1:numel(data)
    if data(i).color > 0
        y = y + sum(data(i).values);
    end
end

The expression coder.varsize('data(:).values') defines the field values inside each element of matrix data to be variable-size.

Here are other examples:

  • coder.varsize('data.A(:).B')

    In this example, data is a scalar variable that contains matrix A. Each element of matrix A contains a variable-size field B.

  • coder.varsize('data(:).A(:).B')

    This expression defines field B inside each element of matrix A inside each element of matrix data to be variable-size.

See Also

Related Topics